By Ben Wilson, Principal Consultant at phData.
The Fifth Job, part 1 of 6.
[DIAGRAM: plain white, five labelled boxes in a row (find, sell, support, admin, deliver); the first four shaded grey and labelled "agent-shaped", the fifth outlined in ink and labelled "the business"; no numbers in the handwriting face.]
Ask an AI agent to run a furniture business and it will happily do the books. It will not build the table.
That gap is not a flaw to wait out. It tells you where agents go.
Five jobs, 2,500 businesses
Most small businesses do five jobs. They find customers, sell to them, deliver the product or service, support customers after the sale, and run the admin: invoices, scheduling, books.
I wanted to know how much of each job an artificial intelligence (AI) agent can carry today. So I had a model score 2,500 businesses from Starter Story, a public library of founder interviews. For each business, the model read a 300-character public description and scored each job from 0 to 3. A 0 means an agent cannot do it, 1 means it assists, 2 means it does most of the job, and 3 means it runs the job. I call a job agent-shaped when it scores 2 or more.
The model also sorted the businesses into 12 types, from software to local services. Two patterns held across them. Admin is the most agent-shaped job in eleven of the twelve; the exception is consumer apps, where delivery is software. Deliver is the least agent-shaped job in nine.
Three businesses, one shape
Frailey Woodworks makes custom furniture: $30k a month, eight people. Four of its five jobs score as agent-shaped. Deliver scores 0. The delivery is the furniture.
Furry Fellas Pet Service does pet sitting and dog walking: $42k a month, nine people. Four of five again. An agent can book the visit and send the invoice. Someone still has to walk the dog.
LogoClothz sells custom-printed tablecloths: $32k a month, five people. Same four, same zero. The logo has to come out right on the cloth.
Look at what the four scores are. Each one is a 2, not a 3: the model says an agent does most of each job, and a person still checks the work. The zero is different. It is not a hard task scored low. It is a different kind of job: hands, a workshop, a leash.
These three are not outliers. Across 142 local service businesses in the data, deliver averages 0.03 of 3. Across 79 handmade businesses it averages 0.13, and across 85 operations and supply businesses, 0.38. Admin in the same three types averages between 1.95 and 2.05. The shape is the same at scale: the paperwork is agent-shaped, and the product is not.
[BEN: one or two sentences from your own non-client life. A small business you know or use (a contractor, a vet, a shop) where the paperwork could run itself but the thing you pay for could not. Nothing from a client or employer engagement.]
The fifth job is the business
The easy reading is that four of five jobs automate, so 80% of the business automates. That gets it backward.
The fifth job is not leftover work. It is why the customer pays. Nobody buys from Frailey Woodworks because the invoices go out on time. They buy the furniture. A dog owner pays Furry Fellas for a person they trust with a key to the house.
For these three businesses, the fifth job is physical. For others it is taste, trust, or the founder's own judgment. An agency sells judgment. A marketplace sells trust. Posts 3 to 5 in [[agentify-3000-series|this series]] take those one at a time.
So the question changes. It stops being "can agents run my business?" It becomes "which jobs around my business should agents carry, so people get more hours for the fifth?" For a furniture maker, that is time back at the bench.
Try it on your business
- Score your five jobs. Use the same scale: 0 (an agent cannot), 1 (assists), 2 (does most), 3 (runs it). Score how each job is done today, not how you would like it done.
- Name your fifth job in one sentence. If the sentence names a tool or a process, try again. The fifth job is what the customer pays for. For a dog-walking business, it might read: a person you trust takes care of your dog.
- Hand over the highest score first, and protect the fifth. Track the hours the fifth job gets each week. If agents take the other four and those hours stay flat, nothing changed that the customer can feel.
The bear case
Starter Story is a self-selected sample. Founders report their own wins, and header numbers can be peak, stale, or wrong. Before naming each business here, I had the agent re-check its public page on October 4, 2026. All three passed (receipts below).
The scores are the weaker evidence. A model graded its own capability from a 300-character blurb. A blurb about a woodworking shop doesn't say how custom quotes get made, and an agent may be worse at that than the model thinks. The same agent spot-checked 30 businesses: its business type matched on 28, and all 30 sets of job scores were within one point. That check was not independent. The agent that did the scoring also ran the check.
[BEN: optional. Re-read the 30 spot-check businesses yourself before publishing if you want to say "I checked" in the first person ([[archetypes]]:74). Otherwise keep the paragraph above as written.]
Last, incumbents can adopt agents too. If every pet sitter automates admin, admin stops being an edge. It becomes the floor, and the fifth job is the only place left to compete.
I hold this view with moderate confidence. The pattern repeats across 2,500 businesses, but the scores behind it come from a blurb.
Your turn
What is your fifth job, in one sentence?
The next post argues for admin first. Before you read it: if four of your jobs are agent-shaped, which one would you hand over first?
Receipts
Method (what ran). Starter Story publishes founder interviews with a header showing monthly revenue and team size. 3,261 public stories were deduplicated by website to 2,566 businesses; 66 with suspect revenue figures were dropped, leaving 2,500. A model, run on a fixed prompt that I directed, assigned each business one of 12 types and scored each of the five jobs 0-3 for agent carry today, reading only a 300-character public description. Aggregates by type: admin is the top-scoring job in eleven of twelve types (never below 1.95 of 3); deliver is the lowest in nine (local service 0.03, handmade 0.13, operations and supply 0.38). Source: Starter Story; scores recomputed from the scored dataset on October 4, 2026.
What broke. The spot-check (28 of 30 business types matched, all 30 job-score sets within one point) was run by the same agent that did the scoring, not by an independent reader. Revenue is self-reported, as of each story's publication date. Most full stories are for Starter Story premium members; the gate below used only the free public portion of each page.
Businesses named (name, revenue and team size as shown in each page header; story year in brackets):
| Business | Revenue | Team | Link |
|---|---|---|---|
| Frailey Woodworks | $30k/month | 8 (2 founders, 6 employees) | Starter Story, 2024 |
| Furry Fellas Pet Service | $42k/month | 9 (1 founder, 8 employees) | Starter Story, 2025 |
| LogoClothz | $32k/month | 5 (2 founders, 3 employees) | Starter Story, 2024 |
Each scored find 2, sell 2, deliver 0, support 2, admin 2.
Deliberately left out. StandOut CV (sold April 2026; story body and header disagree).
LinkedIn cut (180 words, link in first comment)
Ask an artificial intelligence (AI) agent to run a furniture business and it will happily do the books. It will not build the table.
Small businesses do five jobs: find customers, sell, deliver, support, admin.
I had a model score 2,500 public businesses from Starter Story on how much of each job an agent can carry today, 0 to 3.
Three of them, same shape:
Frailey Woodworks, $30k a month, eight people. Four of five jobs agent-shaped. The fifth is the furniture. Furry Fellas Pet Service, $42k a month, nine people. Someone still has to walk the dog. LogoClothz, $32k a month, five people. The logo has to come out right on the cloth.
The fifth job isn't leftover work. It's why the customer pays.
So the question isn't "can agents run my business?" It's "which jobs should agents carry, so people get more hours for the fifth?"
Caveat: a model scored itself from short public blurbs. Treat it as a starting frame.
What's your fifth job, in one sentence?
Full write-up, method and bear case in the first comment.
Foil gate
Checked 2026-10-04 17:07-17:08 EDT by fetching each public Starter Story page (curl, browser user agent, no login). Every story page ends with "Full article is available for Starter Story premium members", so each check covers the free public portion only; no paywall was bypassed. Business websites were also fetched to confirm they are live.
| Foil | URL | Founder still runs it | Header revenue vs body | Team in header | Result |
|---|---|---|---|---|---|
| Frailey Woodworks | https://www.starterstory.com/stories/frailey-woodworks | Yes: header names founder James Frailey; public body describes the family running the shop; no sale or closure mentioned. Site fraileywoodworks.com live (HTTP 200). | Header $30K/mo; body states the current year averages $30,000/month in sales. Agree. | 2 founders + 6 employees = 8 | PASS |
| Furry Fellas Pet Service | https://www.starterstory.com/stories/furry-fellas-pet-service-llc-c6b54520-dece-4ca7-8cf6-f281ae8cb60a (2025 update; earlier 2022 update at /furry-fellas-pet-service-llc) | Yes: header names founder Melodi Landi; 2025 public body describes ongoing operations (consolidated to a smaller radius, adding an e-book and coaching for pet sitters); no sale or closure. Site furryfellaspetsitting.com live (HTTP 200). | Header $42K/mo; headline says $42K/month; public body gives no other figure. No disagreement. (Older stories: $54K/month in 2020; the 2022 update mentions about 25 part-time staff, not shown as a figure in the 2025 header.) | 1 founder + 8 employees = 9 | PASS (note: team of 9 likely counts full-time equivalents, not the part-time roster) |
| LogoClothz | https://www.starterstory.com/stories/logoclothz-385cf77d-944b-4a7e-a2ac-9aa5cc3c5568 (2024 update; 2022 update at /logoclothz shows the same header) | Yes: header names founders Paul Bucklin and Ryan Cunningham; public body describes current operations and expanded product line; no sale or closure. Site logoclothz.com live, copyright 2019-2026. | Header $32K/mo; headline says doubled to $32K/month; public body gives no other figure. No disagreement. | 2 founders + 3 employees = 5 | PASS |
Dataset cross-check (data/businesses-scored.csv): all three rows match the headers above (revenue, team) and score find 2, sell 2, deliver 0, support 2, admin 2. Aggregate claims recomputed from the CSV on 2026-10-04: 2,566 rows, 66 revenue-suspect, 2,500 clean; admin top job in 11 of 12 types (consumer_app top is deliver 2.34 vs admin 2.25), admin minimum 1.95 (ops); deliver lowest in 9 of 12 (not saas, consumer_app, media). The outline's check markers on n, "11 of 12" and "9" are resolved.
No swap needed.